Training and Predicting a 3T Data Set (splice site) on a Machine with 120G RAM
Huei-Shin Chen and Meng-Yuan Yang 2015-07-24
1 Data Training with LIBLINEAR-CDBLOCK
With the file being 2.7TB in size, which is usually larger than the RAM of most machines, typical clas- sifiers such as LIBLINEAR (Fan et al., 2008) cannot be used. The reason is that they store the whole data into the memory for training. To solve this issue, we consider LIBLINEAR-CDBLOCK avail- able at http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/cdblock/. LIBLINEAR-CDBLOCK (Yu et al., 2012) first splits the training data into pieces so that each can be stored into the memory.
Its algorithm requires only one piece of data in the main memory at a time.
1.1 Data Uncompression
The following command uncompresses the training set.
$ unxz splice site.xz
It took about 18.5 hours.
To see if the file is correctly downloaded and uncompressed, you can check the md5sum.
$ md5sum splice site
It took about 7 hours.
If the md5sum is df3bd1b65b9df5776907721dff4fdb4e, the file is correctly uncompressed.
1.2 Data Split
We prepared one machine with around 120G memory for training, so we decided to split the data into 40 pieces.
$ ./blocksplit -m 40 splice site
The running time is around 60 hours.
1
1.3 Training
The following command solves the dual problem of L1-loss support vector machines.
$ ./blocktrain -s 3 splice site.40 splice.s3.model
The training process took about 35 hours (127890 seconds) for finishing 5 outer iterations.
The following command solves the dual problem of L2-regularized logistic regression.
$ ./blocktrain -s 7 splice site.40 splice.s7.model
The training process took about 15 hours (55359 seconds) for finishing 1 outer iteration.
1.4 Prediction
Later we follow Sonnenburg and Franc (2010) to use AUPRC as the evaluation criterion. The calcu- lation of AUPRC requires decision values, so we modify the code to output them. To calculate the AUPRC, decision values and true labels should be printed out during prediction. predict.c needs to be modified for this.
In predict.c change the following lines to print decision values:
predict label = predict(model , x);
fprintf(output,‘‘%g\n’’,predict label);
into:
double dec values;
predict label = predict values(model , x, &dec values);
fprintf(output,‘‘%g %g\n’’, target label, dec values);
The following command predicts instances in the test file splice site.t, and outputs true labels and predicted decision values.
$ ./predict splice site.t splice.s3.model s3 out
$ ./predict splice site.t splice.s7.model s7 out
It took around 45 minutes.
2
1.5 Results
We use Matlab Statistics to calculate AUPRC
>> load s3 out
>> [Xpr,Ypr,Tpr,AUCpr] = perfcurve(s3 out(:,1), s3 out(:,2), 1, ‘xCrit’,
‘reca’, ‘yCrit’, ‘prec’); AUCpr The AUPRC values are
L1-loss support vector machines (dual): 0.5773 L2-regularized logistic regression (dual): 0.5772
These values are similar to those in Sonnenburg and Franc (2010).
References
R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. LIBLINEAR: a library for large linear classification. Journal of Machine Learning Research, 9:1871–1874, 2008. URL http:
//www.csie.ntu.edu.tw/~cjlin/papers/liblinear.pdf.
S. Sonnenburg and V. Franc. COFFIN : A computational framework for linear SVMs. In Proceedings of the Twenty Seventh International Conference on Machine Learning (ICML), pages 999–1006, 2010.
H.-F. Yu, C.-J. Hsieh, K.-W. Chang, and C.-J. Lin. Large linear classification when data cannot fit in memory. ACM Transactions on Knowledge Discovery from Data, 5(4):23:1–23:23, February 2012. URL http://www.csie.ntu.edu.tw/~cjlin/papers/kdd_disk_decomposition.pdf.
3
